Lymph node detection in CT scans using modified U-Net with residual learning and 3D deep network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate lymph node (LN) detection is vital for cancer diagnosis and treatment planning.
• Challenges in CT scans include low contrast, varied LN appearance, and sparse distribution, leading to difficult detection and high false positives.
• Manual LN examination is time-consuming and prone to errors, potentially misdirecting clinical focus.

Purpose of the Study:

• To develop an automated framework for lymph node detection in CT images.
• To enhance detection accuracy and significantly reduce false positives.
• To provide a more efficient and reliable tool for oncological investigations.

Main Methods:

• A two-stage deep learning approach was employed: candidate generation and false positive reduction.
• The first stage utilized a modified U-Net with ResNet for high-sensitivity candidate LN identification.
• The second stage employed a 3D convolutional neural network (CNN) for robust false positive reduction.

Main Results:

• The framework achieved 87% sensitivity with 2.75 false positives per volume (FP/vol.) on mediastinal LN datasets.
• For abdominal LN datasets, 79% sensitivity was obtained with 1.74 FP/vol.
• The method demonstrated competitive sensitivity compared to state-of-the-art approaches with a notable reduction in false positives.

Conclusions:

• An automated framework for lymph node detection was successfully developed using deep learning (U-Net with ResNet and 3D CNNs).
• The approach effectively balances high sensitivity with a low false positive rate.
• This automated system can aid clinicians by improving detection accuracy and potentially preventing unnecessary treatments.